The rapid pace at which AI development is evolving is directly proportional to the concerns it is raising. Fear over the unchecked advance of this technology, while there is still no clear regulatory framework around it (covering everything from the limits of its capabilities to ethical, social, labor, and economic impact issues), is no longer limited to the general public; it has reached industry leaders as well.
At the end of August, Microsoft founder, billionaire, and philanthropist Bill Gates published an essay on his personal blog in which he questioned how the transition to the AI era is being handled and said he did not see “signs that leaders, experts, and communities are addressing AI challenges adequately.” This weekend, it was Dario Amodei, CEO of Anthropic, who stepped in, stating that “we must slow the pace at which we improve the capabilities of AI models” in a lengthy reflection published on his blog.
Amodei put forward a series of proposals, including the involvement of “embedded evaluators” from external and independent organizations, the establishment of “democratic coordination” among companies through shared standards and development limits, and the implementation of “global coordination” through which both democratic and authoritarian countries would forge alliances to verify compliance.
For his part, Sam Altman, CEO of OpenAI, agreed with Amodei and announced that his company is committing “to having independent evaluators with access similar to that of employees”.
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we’ve had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We’ll have more to share soon. https://t.co/1YhhIybZX7 — Sam Altman (@sama) September 12, 2026
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we’ve had at OpenAI in recent weeks.
Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We’ll have more to share soon. https://t.co/1YhhIybZX7
— Sam Altman (@sama) September 12, 2026
While these statements of intent offer a small glimmer of hope in the face of an uncertain future in which, as Gates put it, “AI will either be the greatest equalizer ever invented, or the worst source of injustice,” it is worth taking a closer look at the current landscape.
To do so, we have chosen to rely on data from Artificial Analysis, one of the leading independent platforms for analyzing and comparing AI models, and more specifically on its benchmark, the Intelligence Index v4.3.
Source: Artificial Analysis
As we can see, Anthropic and OpenAI top this ranking (with both companies placing several models in strong positions). Readers can draw their own conclusions as to whether being a market leader makes it easier to call for a slowdown in development, or quite the opposite.
Both Claude Fable 5.1 and GPT-6 Astra score 53 on the index. Both models were launched in early September and generated strong audience interest thanks to their improved reasoning capabilities and their ability to execute complex tasks through agents.
That said, a deeper look at Artificial Analysis’s benchmarks shows that the two AIs are tied in performance and share the same pricing ($10 per million input tokens and $50 per million output tokens), as well as a below-average response speed (65.1 output tokens per second for Fable 5.1 vs. 59.8 for Astra). The most notable difference between them is that Fable 5.1 ultimately ends up being more expensive to use than GPT-6 Astra because of its response style. While GPT-6 Astra is efficient, generating just 60M tokens ($5,324.10 total), Fable 5.1 is more long-winded (190M tokens), pushing its evaluation cost up to $13,128.86.
The third- and fourth-ranked models are also from Anthropic: Claude Opus 5 (51) and Claude Fable 5 (50). The former is an intelligent and proactive model designed for everyday use, with capabilities close to those of Claude Fable 5, but at half the price. Claude Fable 5, meanwhile, delivers strong performance in software engineering, knowledge management, computer vision, and scientific research, among other areas.
It is not until fifth place that we finally find an AI model from a different company, though still a very familiar one. Meta’s Muse Spark 1.3 reaches a score of 48 on the intelligence index. According to the benchmark data, this AI stands out for its high speed (228 output tokens per second), competitive pricing, and multimodal support (text, image, video). Its only drawback is its verbosity, generating 170 million tokens in its evaluation, compared with the median of 90 million.
It is striking that the highest-ranked Google model (Gemini 3.8 Flash) comes in only at 12th place. With a score of 41, this AI is very fast (277.5 tokens per second), affordable, and versatile, since it supports text, image, audio, and video. Although Artificial Analysis’s evaluation cost just $1,622.73 thanks to its favorable pricing, it is also quite “talkative”, having generated 170 million tokens as well.
If we look at the top 10 positions in the ranking, we find that only two of the listed AIs do not come from the United States (GLM-5.3 and Kimi K3, both of Chinese origin). However, if we widen the lens and take the 25 listed models into account, a broader variety of origins begins to emerge.
Of the 25 AIs, 15 are from the United States, 8 from China, one from the United Arab Emirates (K2 Horizon 375B A23B), and one from France (Mistral Medium 3.5). It is clear which two countries are leading the way in the development of this technology. Specifically, the Chinese models are:
GLM-5.3 ranks among the leading intelligence models with a score of 45 on Artificial Analysis’s intelligence index. It is a text-only model offering a 1 million token context window. However, it is expensive compared with other similar models, in addition to being somewhat slow (72 tokens per second) and very “talkative” (it generated 210 million tokens during testing).
As for Kimi K3, this AI scored 44 on the index. The model supports text and image input with a 1 million token context window. However, it is an open-weight model that is quite slow (37 tokens per second) and expensive.
As you can see, the generative AI sector is broad, and it includes a wide range of players, not just OpenAI, Anthropic, or Google. This means that, to ensure development that is safe, ethical, and respectful of both the environment and humanity, robust regulation is needed that matches the current and future challenges posed by the transition to the AI era. An era we have already entered, and whose effects we are already feeling (socially, professionally, environmentally, and educationally).
It is not enough to leave control and the setting of limits in the hands of the developers themselves, because we run the risk of their interests outweighing the common good.
Photo: generated with ChatGPT Image 2.5
Your email address will not be published. Required fields are marked *
Δ